Enhanced dry SO₂ capture estimation using Python-driven computational frameworks with hyperparameter tuning and data augmentation

Dry flue gas desulfurization is an invaluable technique to curb sulfur dioxide. In this study, an in-depth comparison of regression models for predicting sulfur dioxide removal was performed. The data-driven models executed were multilayer perceptron, support vector regressor, random forest, categorical boosting, and light gradient boosting machine. The limited experimental samples were magnified to 342 datasets using the random interpolation and random scaling augmentation procedures and analyzed using the empirical cumulative distribution function and box plots. Model training incorporated grid search with cross-validation to identify the optimal hyperparameter sets. The practicality of the resultant customized models was quantified by leveraging the coefficient of determination, mean squared error and root mean square error. The model complexity arising from hyperparameter configurations was appraised based on the Bayesian information criterion and the Akaike information criterion. SHapley Additive exPlanations was essential for comprehending the prediction mechanism through feature significance and the impact of varying feature thresholds on the predicted output. Results obtained evidence that random forest obtained the strongest accuracy, and generalizability from the high coefficient of determination, and lowest error scores. In addition, it was the least convoluted algorithm according to Akaike information criterion and Bayesian information criterion assessments. The SHapley Additive exPlanations analysis revealed that each model interacts with the metadata features uniquely during training, contributing to a varied selection of dominant factors. This paper endorses the technical implementation of machine learning in dry sulfation processes and provides insights into how computer systems perceive data characteristics and make forecasts. • The application of data augmentation to improve data quality • Execution of layered and ensembled algorithms for SO 2 prediction • Analyzing model complexity and interpreting computation processes